Clinical decision support systems for intensive care units: using artificial neural networks

Which model is superior in predicting ICU survival: artificial intelligence versus conventional approaches

Background

A disease severity classification system is widely used to predict the survival of patients admitted to the intensive care unit with different diagnoses. In the present study, conventional severity classification systems were compared with artificial intelligence predictive models (Artificial Neural Network and Decision Tree) in terms of the prediction of the survival rate of the patients admitted to the intensive care unit.

Methods

This retrospective cohort study was performed on the data of the patients admitted to the ICU of Ghaemshahr’s Razi Teaching Care Center from March 20th, 2017, to September 22nd, 2019. The required data for calculating conventional severity classification models (SOFA, SAPS II, APACHE II, and APACHE IV) were collected from the patients’ medical records. Subsequently, the score of each model was calculated. Artificial intelligence predictive models (Artificial Neural Network and Decision Tree) were developed in the next step. Lastly, the performance of each model in predicting the survival of the patients admitted to the intensive care unit was evaluated using the criteria of sensitivity, specificity, accuracy, F-measure, and area under the ROC curve. Also, each model was validated externally. The R program, version 4.1, was used to create the artificial intelligence models, and SPSS Statistics Software, version 21, was utilized to perform statistical analysis.

Results

The area under the ROC curve of SOFA, SAPS II, APACHE II, APACHE IV, multilayer perceptron artificial neural network, and CART decision tree were 76.0, 77.1, 80.3, 78.5, 84.1, and 80.0, respectively.

Conclusion

The results showed that although the APACHE II model had better results than other conventional models in predicting the survival rate of the patients admitted to the intensive care unit, the other conventional models provided acceptable results too. Moreover, the findings showed that the artificial neural network model had the best performance among all the studied models, indicating the discrimination power of this model in predicting patient survival compared to the other models.

Classifying muscle parameters with artificial neural networks and simulated lateral pinch data

Objective

Hill-type muscle models are widely employed in simulations of human movement. Yet, the parameters underlying these models are difficult or impossible to measure in vivo. Prior studies demonstrate that Hill-type muscle parameters are encoded within dynamometric data. But, a generalizable approach for estimating these parameters from dynamometric data has not been realized. We aimed to leverage musculoskeletal models and artificial neural networks to classify one Hill-type muscle parameter (maximum isometric force) from easily measurable dynamometric data (simulated lateral pinch force). We tested two neural networks (feedforward and long short-term memory) to identify if accounting for dynamic behavior improved accuracy.

Methods

We generated four datasets via forward dynamics, each with increasing complexity from adjustments to more muscles. Simulations were grouped and evaluated to show how varying the maximum isometric force of thumb muscles affects lateral pinch force. Both neural networks classified these groups from lateral pinch force alone.

Results

Both neural networks achieved accuracies above 80% for datasets which varied only the flexor pollicis longus and/or the abductor pollicis longus. The inclusion of muscles with redundant functions dropped model accuracies to below 30%. While both neural networks were consistently more accurate than random guess, the long short-term memory model was not consistently more accurate than the feedforward model.

Conclusion

Our investigations demonstrate that artificial neural networks provide an inexpensive, data-driven approach for approximating Hill-type muscle-tendon parameters from easily measurable data. However, muscles of redundant function or of little impact to force production make parameter classification more challenging.

AI in the Intensive Care Unit: Up-to-Date Review

AI is the latest technologic trend that likely will have a huge impact in medicine. AI's potential lies in its ability to process large volumes of data and perform complex pattern analyses. The ICU is an area of medicine that is particularly conducive to AI applications. Much AI ICU research currently is focused on improving high volumes of data on high-risk patients and making clinical workflow more efficient. Emerging topics of AI medicine in the ICU include AI sensors, sepsis prediction, AI in the NICU or SICU, and the legal role of AI in medicine. This review will cover the current applications of AI medicine in the ICU, potential pitfalls, and other AI medicine-related topics relevant for the ICU.

Estimation of Neonatal Intestinal Perforation Associated with Necrotizing Enterocolitis by Machine Learning Reveals New Key Factors

Intestinal perforation (IP) associated with necrotizing enterocolitis (NEC) is one of the leading causes of mortality in premature neonates; with major nutritional and neurodevelopmental sequelae. Since predicting which neonates will develop perforation is still challenging; clinicians might benefit considerably with an early diagnosis tool and the identification of critical factors. The aim of this study was to forecast IP related to NEC and to investigate the predictive quality of variables; based on a machine learning-based technique. The Back-propagation neural network was used to train and test the models with a dataset constructed from medical records of the NICU; with birth and hospitalization maternal and neonatal clinical; feeding and laboratory parameters; as input variables. The outcome of the models was diagnosis: (1) IP associated with NEC; (2) NEC or (3) control (neither IP nor NEC). Models accurately estimated IP with good performances; the regression coefficients between the experimental and predicted data were R² > 0.97. Critical variables for IP prediction were identified: neonatal platelets and neutrophils; orotracheal intubation; birth weight; sex; arterial blood gas parameters (pCO₂ and HCO₃); gestational age; use of fortifier; patent ductus arteriosus; maternal age and maternal morbidity. These models may allow quality improvement in medical practice.

Cross-Disciplinary Consultancy to Enhance Predictions of Asthma Exacerbation Risk in Boston

This paper continues an initiative conducted by the International Society for Disease Surveillance with funding from the Defense Threat Reduction Agency to connect near-term analytical needs of public health practice with technical expertise from the global research community. The goal is to enhance investigation capabilities of day-to-day population health monitors. A prior paper described the formation of consultancies for requirements analysis and dialogue regarding costs and benefits of sustainable analytic tools. Each funded consultancy targets a use case of near-term concern to practitioners. The consultancy featured here focused on improving predictions of asthma exacerbation risk in demographic and geographic subdivisions of the city of Boston, Massachusetts, USA based on the combination of known risk factors for which evidence is routinely available. A cross-disciplinary group of 28 stakeholders attended the consultancy on March 30-31, 2016 at the Boston Public Health Commission. Known asthma exacerbation risk factors are upper respiratory virus transmission, particularly in school-age children, harsh or extreme weather conditions, and poor air quality. Meteorological subject matter experts described availability and usage of data sources representing these risk factors. Modelers presented multiple analytic approaches including mechanistic models, machine learning approaches, simulation techniques, and hybrids. Health department staff and local partners discussed surveillance operations, constraints, and operational system requirements. Attendees valued the direct exchange of information among public health practitioners, system designers, and modelers. Discussion finalized design of an 8-year de-identified dataset of Boston ED patient records for modeling partners who sign a standard data use agreement.

Boundedness and convergence analysis of weight elimination for cyclic training of neural networks

Weight elimination offers a simple and efficient improvement of training algorithm of feedforward neural networks. It is a general regularization technique in terms of the flexible scaling parameters. Actually, the weight elimination technique also contains the weight decay regularization for a large scaling parameter. Many applications of this technique and its improvements have been reported. However, there is little research concentrated on its convergence behavior. In this paper, we theoretically analyze the weight elimination for cyclic learning method and determine the conditions for the uniform boundedness of weight sequence, and weak and strong convergence. Based on the assumed network parameters, the optimal choice for the scaling parameter can also be determined. Moreover, two illustrative simulations have been done to support the theoretical explorations as well.

Diagnosis system for hepatocellular carcinoma based on fractal dimension of morphometric elements integrated in an artificial neural network

BACKGROUND AND AIMS

Hepatocellular carcinoma (HCC) remains a leading cause of death by cancer worldwide. Computerized diagnosis systems relying on novel imaging markers gained significant importance in recent years. Our aim was to integrate a novel morphometric measurement--the fractal dimension (FD)--into an artificial neural network (ANN) designed to diagnose HCC.

MATERIAL AND METHODS

The study included 21 HCC and 28 liver metastases (LM) patients scheduled for surgery. We performed hematoxylin staining for cell nuclei and CD31/34 immunostaining for vascular elements. We captured digital images and used an in-house application to segment elements of interest; FDs were calculated and fed to an ANN which classified them as malignant or benign, further identifying HCC and LM cases.

RESULTS

User intervention corrected segmentation errors and fractal dimensions were calculated. ANNs correctly classified 947/1050 HCC images (90.2%), 1021/1050 normal tissue images (97.23%), 1215/1400 LM (86.78%), and 1372/1400 normal tissues (98%). We obtained excellent interobserver agreement between human operators and the system.

CONCLUSION

We successfully implemented FD as a morphometric marker in a decision system, an ensemble of ANNs designed to differentiate histological images of normal parenchyma from malignancy and classify HCCs and LMs.

A clinical decision support system for femoral peripheral arterial disease treatment

One of the major challenges of providing reliable healthcare services is to diagnose and treat diseases in an accurate and timely manner. Recently, many researchers have successfully used artificial neural networks as a diagnostic assessment tool. In this study, the validation of such an assessment tool has been developed for treatment of the femoral peripheral arterial disease using a radial basis function neural network (RBFNN). A data set for training the RBFNN has been prepared by analyzing records of patients who had been treated by the thoracic and cardiovascular surgery clinic of a university hospital. The data set includes 186 patient records having 16 characteristic features associated with a binary treatment decision, namely, being a medical or a surgical one. K-means clustering algorithm has been used to determine the parameters of radial basis functions and the number of hidden nodes of the RBFNN is determined experimentally. For performance evaluation, the proposed RBFNN was compared to three different multilayer perceptron models having Pareto optimal hidden layer combinations using various performance indicators. Results of comparison indicate that the RBFNN can be used as an effective assessment tool for femoral peripheral arterial disease treatment.

Contrast-enhanced ultrasonography parameters in neural network diagnosis of liver tumors

AIM: To study the role of time-intensity curve (TIC) analysis parameters in a complex system of neural networks designed to classify liver tumors.

METHODS: We prospectively included 112 patients with hepatocellular carcinoma (HCC) (n = 41), hypervascular (n = 20) and hypovascular (n = 12) liver metastases, hepatic hemangiomas (n = 16) or focal fatty changes (n = 23) who underwent contrast-enhanced ultrasonography in the Research Center of Gastroenterology and Hepatology, Craiova, Romania. We recorded full length movies of all contrast uptake phases and post-processed them offline by selecting two areas of interest (one for the tumor and one for the healthy surrounding parenchyma) and consecutive TIC analysis. The difference in maximum intensities, the time to reaching them and the aspect of the late/portal phase, as quantified by the neural network and a ratio between median intensities of the central and peripheral areas were analyzed by a feed forward back propagation multi-layer neural network which was trained to classify data into five distinct classes, corresponding to each type of liver lesion.

RESULTS: The neural network had 94.45% training accuracy (95% CI: 89.31%-97.21%) and 87.12% testing accuracy (95% CI: 86.83%-93.17%). The automatic classification process registered 93.2% sensitivity, 89.7% specificity, 94.42% positive predictive value and 87.57% negative predictive value. The artificial neural networks (ANN) incorrectly classified as hemangyomas three HCC cases and two hypervascular metastases, while in turn misclassifying four liver hemangyomas as HCC (one case) and hypervascular metastases (three cases). Comparatively, human interpretation of TICs showed 94.1% sensitivity, 90.7% specificity, 95.11% positive predictive value and 88.89% negative predictive value. The accuracy and specificity of the ANN diagnosis system was similar to that of human interpretation of the TICs (P = 0.225 and P = 0.451, respectively). Hepatocellular carcinoma cases showed contrast uptake during the arterial phase followed by wash-out in the portal and first seconds of the late phases. For the hypovascular metastases did not show significant contrast uptake during the arterial phase, which resulted in negative differences between the maximum intensities. We registered wash-out in the late phase for most of the hypervascular metastases. Liver hemangiomas had contrast uptake in the arterial phase without agent wash-out in the portal-late phases. The focal fatty changes did not show any differences from surrounding liver parenchyma, resulting in similar TIC patterns and extracted parameters.

CONCLUSION: Neural network analysis of contrast-enhanced ultrasonography - obtained TICs seems a promising field of development for future techniques, providing fast and reliable diagnostic aid for the clinician.

A Comparison of Intensive Care Unit Mortality Prediction Models through the Use of Data Mining Techniques

OBJECTIVES

The intensive care environment generates a wealth of critical care data suited to developing a well-calibrated prediction tool. This study was done to develop an intensive care unit (ICU) mortality prediction model built on University of Kentucky Hospital (UKH)'s data and to assess whether the performance of various data mining techniques, such as the artificial neural network (ANN), support vector machine (SVM) and decision trees (DT), outperform the conventional logistic regression (LR) statistical model.

METHODS

The models were built on ICU data collected regarding 38,474 admissions to the UKH between January 1998 and September 2007. The first 24 hours of the ICU admission data were used, including patient demographics, admission information, physiology data, chronic health items, and outcome information.

RESULTS

Only 15 study variables were identified as significant for inclusion in the model development. The DT algorithm slightly outperformed (AUC, 0.892) the other data mining techniques, followed by the ANN (AUC, 0.874), and SVM (AUC, 0.876), compared to that of the APACHE III performance (AUC, 0.871).

CONCLUSIONS

With fewer variables needed, the machine learning algorithms that we developed were proven to be as good as the conventional APACHE III prediction.

Seven-days-ahead forecasting of childhood asthma admissions using artificial neural networks in Athens, Greece

Artificial Neural Network (ANN) models were developed and applied in order to predict the total weekly number of Childhood Asthma Admission (CAA) at the greater Athens area (GAA) in Greece. Hourly meteorological data from the National Observatory of Athens and ambient air pollution data from seven different areas within the GAA for the period 2001-2004 were used. Asthma admissions for the same period were obtained from hospital registries of the three main Children's Hospitals of Athens. Three different ANN models were developed and trained in order to forecast the CAA for the subgroups of 0-4, 5-14-year olds, and for the whole study population. The results of this work have shown that ANNs could give an adequate forecast of the total weekly number of CAA in relation to the bioclimatic and air pollution conditions. The forecasted numbers are in very good agreement with the observed real total weekly numbers of CAA.

New developmental stages for common marmosets (Callithrix jacchus) using mass and age variables obtained by K-means algorithm and self-organizing maps (SOM)

This study proposes new developmental stages for Callithrix jacchus, using K-Means algorithm and an artificial neural network-self-organising maps (SOM) as computational tools, based on weight and age. Eight developmental stages are proposed: Infant I, II and III, Juvenile I and II, Sub adult, Young adult and Older adult. This classification is consistent with the first appearance of several behavioural and physiological characteristics and thus may have generality in defining critical developmental periods. It also reveals differences in male and female development and establishes a stage for the onset of the final adult life cycle. This classification is also important to understanding the biology of the ontogenetic development of common marmosets, providing new insights for the management and care of captive animals and improving age estimate indicators when specimens are captured in long term monitoring of free ranging groups.

Complimentary artificial neural network approaches for prediction of events in the neonatal intensive care unit

In the neonatal intensive care unit, the early and accurate prediction of mortality, length of stay and duration of ventilation can improve decision making. For physiological events, non-linear prediction models generally out-perform statistical-based approaches, as was confirmed in these experiments. For three medical outcomes, the maximum-likelihood (ML) approximation was used in conjunction with a gradient descent artificial neural network (ANN) prototype to create models with risk estimation ranges. The ML ANN showed that the ML estimation function was successful at creating variable sensitivity models for three important outcomes. The flexibility of the ML ANN in terms of output values differentiates it from the more traditional ANN.

A hybrid hierarchical decision support system for cardiac surgical intensive care patients. Part I: Physiological modelling and decision support system design

OBJECTIVE

To develop a clinical decision support system (CDSS) that models the different levels of the clinician's decision-making strategies when controlling post cardiac surgery patients weaned from cardio pulmonary bypass.

METHODS

A clinical trial was conducted to define and elucidate an expert anesthetists' decision pathway utilised in controlling this patient population. This data and derived knowledge were used to elicit a decision-making model. The structural framework of the decision-making model is hierarchical, clearly defined, and dynamic. The decision levels are linked to five important components of the cardiovascular physiology in turn, i.e. the systolic blood pressure (SBP), central venous pressure (CVP), systemic vascular resistance (SVR), cardiac output (CO), and heart rate (HR). Progress down the hierarchy is dependent upon the normalisation of each physiological parameter to a value pre-selected by the clinician via fluid, chronotropes or inotropes. Since interventions at each and every level cause changes and disturbances in the other components, the proposed decision support model continuously refers back decision outcomes back to the SBP which is considered to be the overriding supervisory safety component in this hierarchical decision structure. The decision model was then translated into a computerised decision support system prototype and comprehensively tested on a physiological model of the human cardiovascular system. This model was able to reproduce conditions experienced by post-operative cardiac surgery patients including hypertension, hypovolemia, vasodilation and the systemic inflammatory response syndrome (SIRS).

RESULTS

In all the simulated patients scenarios considered the CDSS was able to initiate similar therapeutic interventions to that of the expert, and as a result, was also able to control the hemodynamic parameters to the prescribed target values.

Validation study of artificial neural network models for prediction of methicillin-resistant Staphylococcus aureus carriage

OBJECTIVE

Use of active surveillance cultures for methicillin-resistant Staphylococcus aureus (MRSA) for all patients admitted to the intensive care unit has been shown to reduce nosocomial transmission. However, the cost-effectiveness and the utility of implementing use of active surveillance cultures nationwide remain controversial. We sought to develop an artificial neural network (ANN) model that would predict the likelihood of MRSA colonization.

SETTING

Two acute care hospitals, one in Pittsburgh (hospital A) and one in Kaohsiung, Taiwan (hospital B).

METHODS

Nasal cultures were performed for all patients admitted to the hospitals. A total of 46 potential risk factors in hospital A and 86 potential risk factors in hospital B associated with MRSA colonization were assessed. Culture results were obtained; 75% of the data were used for training our ANN model, and the remaining 25% were used for validating our ANN model. The culture results were the "gold standard" for determining the accuracy of the model predictions.

RESULTS

The ANN model predictions were accurate 95.2% of the time for hospital A (sensitivity, 94.3%; specificity, 96.0%) and 94.2% of the time for hospital B (sensitivity, 96.6%; specificity, 91.8%), integrating all potential risk factors into the model. Only 17 potential risk factors were needed for the hospital A ANN model (accuracy, 90.9%; sensitivity, 98.5%; specificity, 83.4%), and only 20 potential risk factors were needed for the hospital B ANN model (accuracy, 90.5%; sensitivity, 96.6%; specificity, 84.3%), if the minimal risk factor method was used. Cross-validation analysis showed an average accuracy of 85.6% (sensitivity, 91.3%; specificity, 80.0%).

CONCLUSION

Our ANN model can be used to predict with an accuracy of more than 90% which patients carry MRSA. The false-negative rates were significantly lower than the false-positive rates in the ANN predictions, which can serve as a safety buffer in case of patient misclassification.

A novel approach for prediction of tacrolimus blood concentration in liver transplantation patients in the intensive care unit through support vector regression

Introduction

Tacrolimus is an important immunosuppressive drug for organ transplantation patients. It has a narrow therapeutic range, toxic side effects, and a blood concentration with wide intra- and interindividual variability. Hence, it is of the utmost importance to monitor tacrolimus blood concentration, thereby ensuring clinical effect and avoiding toxic side effects. Prediction models for tacrolimus blood concentration can improve clinical care by optimizing monitoring of these concentrations, especially in the initial phase after transplantation during intensive care unit (ICU) stay. This is the first study in the ICU in which support vector machines, as a new data modeling technique, are investigated and tested in their prediction capabilities of tacrolimus blood concentration. Linear support vector regression (SVR) and nonlinear radial basis function (RBF) SVR are compared with multiple linear regression (MLR).

Methods

Tacrolimus blood concentrations, together with 35 other relevant variables from 50 liver transplantation patients, were extracted from our ICU database. This resulted in a dataset of 457 blood samples, on average between 9 and 10 samples per patient, finally resulting in a database of more than 16,000 data values. Nonlinear RBF SVR, linear SVR, and MLR were performed after selection of clinically relevant input variables and model parameters. Differences between observed and predicted tacrolimus blood concentrations were calculated. Prediction accuracy of the three methods was compared after fivefold cross-validation (Friedman test and Wilcoxon signed rank analysis).

Results

Linear SVR and nonlinear RBF SVR had mean absolute differences between observed and predicted tacrolimus blood concentrations of 2.31 ng/ml (standard deviation [SD] 2.47) and 2.38 ng/ml (SD 2.49), respectively. MLR had a mean absolute difference of 2.73 ng/ml (SD 3.79). The difference between linear SVR and MLR was statistically significant (p < 0.001). RBF SVR had the advantage of requiring only 2 input variables to perform this prediction in comparison to 15 and 16 variables needed by linear SVR and MLR, respectively. This is an indication of the superior prediction capability of nonlinear SVR.

Conclusion

Prediction of tacrolimus blood concentration with linear and nonlinear SVR was excellent, and accuracy was superior in comparison with an MLR model.

Imputation of missing values by integrating neural networks and case-based reasoning

Missing values in a medical database present a problem when trying to develop a prediction model for a broad range of patients, if the data are not missing at random. We present a data imputation approach for physiologic parameters that incorporates individualized case information into the imputed values. We replaced missing values in a neonatal intensive care unit (NICU) database with relevant data by integrating aspects of artificial neural networks (ANNs) and case-based reasoning (CBR).

The use of artificial neural networks to stratify the length of stay of cardiac patients based on preoperative and initial postoperative factors

BACKGROUND

The limitations of current prognostic models in identifying postoperative cardiac patients at risk of experiencing morbidity and subsequently an extended intensive care unit length of stay (ICU LOS) is well recognized. This coupled with the desire for risk stratification in order to prioritize medical intervention has lead to the need for the development of a system that can accurately predict individual patient outcome based on both preoperative and immediate postoperative clinical factors. The usefulness of artificial neural networks (ANNs) as an outcome prediction tool in the critical care environment has been previously demonstrated for medical intensive care unit (ICU) patients and it is the aim of this study to apply this methodology to postoperative cardiac patients.

METHODS

A review of contemporary literature revealed 15 preoperative risk factors and 17 operative and postoperative variables that have a determining effect on LOS. An integrated, multi-functional software package was developed to automate the ANN development process. The efficacy of the resultant individual ANNs as well as groupings or ensembles of ANNs were measured by calculating sensitivity and specificity estimates as well as the area under the receiver operating curve (AUC) when the ANN is applied to an independent test dataset.

RESULTS

The individual ANN with the highest discriminating ability produced an AUC of 0.819. The use of the ensembles of networks technique significantly improved the classification accuracy. Consolidating the output of three ANNs improved the AUC to 0.90.

CONCLUSIONS

This study demonstrates the suitability of ANNs, in particular ensembles of ANNs, to outcome prediction tasks in postoperative cardiac patients.

Double-blind evaluation and benchmarking of survival models in a multi-centre study

Accurate modelling of time-to-event data is of particular importance for both exploratory and predictive analysis in cancer, and can have a direct impact on clinical care. This study presents a detailed double-blind evaluation of the accuracy in out-of-sample prediction of mortality from two generic non-linear models, using artificial neural networks benchmarked against a partial logistic spline, log-normal and COX regression models. A data set containing 2880 samples was shared over the Internet using a purpose-built secure environment called GEOCONDA (www.geoconda.com). The evaluation was carried out in three parts. The first was a comparison between the predicted survival estimates for each of the four survival groups defined by the TNM staging system, against the empirical estimates derived by the Kaplan-Meier method. The second approach focused on the accurate prediction of survival over time, quantified with the time dependent C index (C(td)). Finally, calibration plots were obtained over the range of follow-up and tested using a generalization of the Hosmer-Lemeshow test. All models showed satisfactory performance, with values of C(td) of about 0.7. None of the models showed a systematic tendency towards over/under estimation of the observed survival at tau=3 and 5 years. At tau=10 years, all models underestimated the observed survival, except for COX regression which returned an overestimate. The study presents a robust and unbiased benchmarking methodology using a bespoke web facility. It was concluded that powerful, recent flexible modelling algorithms show a comparative predictive performance to that of more established methods from the medical and biological literature, for the reference data set.

Prognosis in critical care

Prognostic risk prediction models have been employed in the intensive care unit (ICU) setting since the 1980s and provide health care providers with important information to help inform decisions related to treatment and prognosis, as well as to compare outcomes across institutions. Prognostic models for critical care are among the most widely utilized and tested predictive models in healthcare. In this article, we review and compare mortality prediction models, including the APACHE (1981), SAPS (1984), APACHE-II (1985), MPM (1987), APACHE-III (1991), SAPS-II (1993), and MPM-II (1993). We emphasize the importance of model calibration in this domain. In addition, we present a brief review of the statistical methodology, multiple logistic regression, which underlies most of the models currently used in critical care.

Predicting high-risk preterm birth using artificial neural networks

A reengineered approach to the early prediction of preterm birth is presented as a complimentary technique to the current procedure of using costly and invasive clinical testing on high-risk maternal populations. Artificial neural networks (ANNs) are employed as a screening tool for preterm birth on a heterogeneous maternal population; risk estimations use obstetrical variables available to physicians before 23 weeks gestation. The objective was to assess if ANNs have a potential use in obstetrical outcome estimations in low-risk maternal populations. The back-propagation feedforward ANN was trained and tested on cases with eight input variables describing the patient's obstetrical history; the output variables were: 1) preterm birth; 2) high-risk preterm birth; and 3) a refined high-risk preterm birth outcome excluding all cases where resuscitation was delivered in the form of free flow oxygen. Artificial training sets were created to increase the distribution of the underrepresented class to 20%. Training on the refined high-risk preterm birth model increased the network's sensitivity to 54.8%, compared to just over 20% for the nonartificially distributed preterm birth model.

Classification of EMG signals using PCA and FFT

In this study, the fast Fourier transform (FFT) analysis was applied to EMG signals recorded from ulnar nerves of 59 patients to interpret data. The data of the patients were diagnosed by the neurologists as 19 patients were normal, 20 patients had neuropathy and 20 patients had myopathy. The amount of FFT coefficients had been reduced by using principal components analysis (PCA). This would facilitate calculation and storage of EMG data. PCA coefficients were applied to multilayer perceptron (MLP) and support vector machine (SVM) and both classified systems of performance values were computed. Consequently, the results show that SVM has high anticipation level in the diagnosis of neuromuscular disorders. It is proved that its test performance is high compared with MLP.

A novel dengue fever (DF) and dengue haemorrhagic fever (DHF) analysis using artificial neural network (ANN)

Dengue fever (DF) is an acute febrile viral disease frequently presented with headache, bone or joint and muscular pains, and rash. A significant percentage of DF patients develop a more severe form of disease, known as dengue haemorrhagic fever (DHF). DHF is the complication of DF. The main pathophysiology of DHF is the development of plasma leakage from the capillary, resulting in haemoconcentration, ascites, and pleural effusion that may lead to shock following defervescence of fever. Therefore, accurate prediction of the day of defervescence of fever is critical for clinician to decide on patient management strategy. To date, no known literature describes of any attempt to predict the day of defervescence of fever in DF patients. This paper describes a non-invasive prediction system for predicting the day of defervescence of fever in dengue patients using artificial neural network. The developed system bases its prediction solely on the clinical symptoms and signs and uses the multilayer feed-forward neural networks (MFNN). The results show that the proposed system is able to predict the day of defervescence in dengue patients with 90% prediction accuracy.

Conceptual Framework of Knowledge Management for Ethical Decision-Making Support in Neonatal Intensive Care

This research is built on the belief that artificial intelligence estimations need to be integrated into clinical social context to create value for health-care decisions. In sophisticated neonatal intensive care units (NICUs), decisions to continue or discontinue aggressive treatment are an integral part of clinical practice. High-quality evidence supports clinical decision-making, and a decision-aid tool based on specific outcome information for individual NICU patients will provide significant support for parents and caregivers in making difficult "ethical" treatment decisions. In our approach, information on a newborn patient's likely outcomes is integrated with the physician's interpretation and parents' perspectives into codified knowledge. Context-sensitive content adaptation delivers personalized and customized information to a variety of users, from physicians to parents. The system provides structuralized knowledge translation and exchange between all participants in the decision, facilitating collaborative decision-making that involves parents at every stage on whether to initiate, continue, limit, or terminate intensive care for their infant.

Use of Support Vector Machines and Neural Network in Diagnosis of Neuromuscular Disorders

In this study the performance of support vector machine (SVM)and back-propagation neural network were applied to analyze the classification of the electromyogram (EMG) signals obtained from normal, neuropathy and myopathy subjects. By using autoregressive (AR) modeling, AR coefficients were obtained from EMG signals. Moreover, the support vector machine and artificial neural network (ANN) were used as base classifiers. The AR coefficients were benefited as inputs for SVM and ANN. Besides, these coefficients were tested both in ANN and SVM. The results show that SVM has high anticipation level in the diagnosis of neuromuscular disorders. It is proved that its test performance is high compared with ANN.

Modelling survival after treatment of intraocular melanoma using artificial neural networks and Bayes theorem

This paper describes the development of an artificial intelligence (AI) system for survival prediction from intraocular melanoma. The system used artificial neural networks (ANNs) with five input parameters: coronal and sagittal tumour location, anterior tumour margin, largest basal tumour diameter and the cell type. After excluding records with missing data, 2331 patients were included in the study. These were split randomly into training and test sets. Date censorship was applied to the records to deal with patients who were lost to follow-up and patients who died from general causes. Bayes theorem was then applied to the ANN output to construct survival probability curves. A validation set with 34 patients unseen to both training and test sets was used to compare the AI system with Cox's regression (CR) and Kaplan-Meier (KM) analyses. Results showed large differences in the mean 5 year survival probability figures when the number of records with matching characteristics was small. However, as the number of matches increased to > 100 the system tended to agree with CR and KM. The validation set was also used to compare the system with a clinical expert in predicting time to metastatic death. The rms error was 3.7 years for the system and 4.3 years for the clinical expert for 15 years survival. For < 10 years survival, these figures were 2.7 and 4.2, respectively. We concluded that the AI system can match if not better the clinical expert's prediction. There were significant differences with CR and KM analyses when the number of records was small, but it was not known which model is more accurate.

Improvement and automation of artificial neural networks to estimate medical outcomes

The lengthy process of manually optimizing a feedforward backpropagation artificial neural network (ANN) provided the incentive to develop an automated system that could fine-tune the network parameters without user supervision. A new stopping criterion was introduced--the logarithmic-sensitivity index--that manages a good balance between sensitivity and specificity of the output classification. The automated network automatically monitored the classification performance to determine when was the best time to stop training-after no improvement in the performance measure (either highest correct classification rate, lowest mean squared error or highest log-sensitivity index value) occurred in the subsequent 500 epochs. Experiments were performed on three medical databases: an adult intensive care unit, a neonatal intensive care unit and a coronary surgery patient database. The optimal network parameter settings found by the automated system were similar to those found manually. The results showed that the automated networks performed equally well or better than the manually optimized ANNs, and the best classification performance was achieved using the log-sensitivity index as a stopping criterion.

Prediction of mortality in an Indian intensive care unit. Comparison between APACHE II and artificial neural networks

OBJECTIVE

To compare hospital outcome prediction using an artificial neural network model, built on an Indian data set, with the APACHE II (Acute Physiology and Chronic Health Evaluation II) logistic regression model.

DESIGN

Analysis of a database containing prospectively collected data.

SETTING

Medical-neurological ICU of a university hospital in Mumbai, India.

SUBJECTS

Two thousand sixty-two consecutive admissions between 1996 and 1998.

INTERVENTIONS

None.

MEASUREMENTS AND RESULTS

The 22 variables used to obtain day-1 APACHE II score and risk of death were recorded. Data from 1,962 patients were used to train the neural network using a back-propagation algorithm. Data from the remaining 1,000 patients were used for testing this model and comparing it with APACHE II. There were 337 deaths in these 1,000 patients; APACHE II predicted 246 deaths while the neural network predicted 336 deaths. Calibration, assessed by the Hosmer-Lemeshow statistic, was better with the neural network (H=22.4) than with APACHE II (H=123.5) and so was discrimination (area under receiver operating characteristic curve =0.87 versus 0.77, p=0.002). Analysis of information gain due to each of the 22 variables revealed that the neural network could predict outcome using only 15 variables. A new model using these 15 variables predicted 335 deaths, had calibration (H=27.7) and discrimination (area under receiver operating characteristic curve =0.88) which was comparable to the 22-variable model (p=0.87) and superior to the APACHE II equation (p<0.001).

CONCLUSION

Artificial neural networks, trained on Indian patient data, used fewer variables and yet outperformed the APACHE II system in predicting hospital outcome.

Weight-elimination neural networks applied to coronary surgery mortality prediction

The objective was to assess the effectiveness of the weight-elimination cost function in improving classification performance of artificial neural networks (ANNs) and to observe how changing the a priori distribution of the training set affects network performance. Backpropagation feedforward ANNs with and without weight-elimination estimated mortality for coronary artery surgery patients. The ANNs were trained and tested on cases with 32 input variables describing the patient's medical history; the output variable was in-hospital mortality (mortality rates: training 3.7%, test 3.8%). Artificial training sets with mortality rates of 20%, 50%, and 80% were created to observe the impact of training with a higher-than-normal prevalence. When the results were averaged, weight-elimination networks achieved higher sensitivity rates than those without weight-elimination. Networks trained on higher-than-normal prevalence achieved higher sensitivity rates at the cost of lower specificity and correct classification. The weight-elimination cost function can improve the classification performance when the network is trained with a higher-than-normal prevalence. A network trained with a moderately high artificial mortality rate (artificial mortality rate of 20%) can improve the sensitivity of the model without significantly affecting other aspects of the model's performance. The ANN mortality model achieved comparable performance as additive and statistical models for coronary surgery mortality estimation in the literature.

Recent innovations in intensive care unit risk-prediction models

During the past 20 years, ICU risk-prediction models have undergone significant development, validation, and refinement. Among the general ICU severity of illness scoring systems, the Acute Physiology and Chronic Health Evaluation (APACHE), Mortality Prediction Model (MPM), and the Simplified Acute Physiology Score (SAPS) have become the most accepted and used. To risk-adjust patients with longer, more severe illnesses like sepsis and acute respiratory distress syndrome, several models of organ dysfunction or failure have become available, including the Multiple Organ Dysfunction Score (MODS), the Sequential Organ Failure Assessment (SOFA), and the Logistic Organ Dysfunction Score (LODS). Recent innovations in risk adjustment include automatic physiology and diagnostic variable retrieval and the use of artificial intelligence. These innovations have the potential of extending the uses of case-mix and severity-of-illness adjustment in the areas of clinical research, patient care, and administration. The challenges facing intensivists in the next few years are to further develop these models so that they can be used throughout the IUC stay to assess quality of care and to extend them to more specific patient groups such as the elderly and patients with chronic ICU courses.